Risk control model training method, risk prediction method, device, and computer equipment

By calculating the correlation coefficients between the feature column vector and the classification identification vector, the feature labels are screened and the risk control model parameters are adjusted, and the accuracy of the Internet product risk control model is low when it is cold-started, and a higher risk control accuracy is achieved.

CN115630318BActive Publication Date: 2024-11-26CLP JINXIN SOFTWARE (SHANGHAI CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211403924.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-11-10
Publication Date
2024-11-26
Estimated Expiration
2042-11-10

AI Technical Summary

Technical Problem

When the risk control model is cold-started, due to the lack of sufficient data, the selection of feature labels is highly subjective and the accuracy of the risk control model is low.

Method used

By calculating the correlation coefficient between the feature column vector and the preset classification identification vector, filter out the feature labels that meet the preset conditions, build an initial risk control model, and adjust the model parameters by exchanging the feature activation set and the feature labels in the non-feature activation set. Until the preset error threshold is reached, the accuracy of the risk control model is improved.

Benefits of technology

It improves the prediction accuracy of the risk control model, reduces manual intervention and subjectivity, and improves the risk control capabilities of Internet products that have not been in operation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115630318B_ABST
    Figure CN115630318B_ABST
Patent Text Reader

Abstract

The present application relates to a training method, a risk prediction method, an apparatus, and a computer device for a risk control model. The method comprises: obtaining a feature column vector corresponding to a feature label according to a plurality of feature labels contained in a risk control training sample set and a feature label value corresponding to the feature label. Determine the correlation coefficient between the feature column vector and a preset classification identification vector, and obtain the correlation coefficient between the feature label and the classification identification vector. Among the feature labels of the risk control training sample, filter the feature labels whose corresponding correlation coefficients meet the preset feature screening conditions, add the filtered feature labels to the feature activation set, and finally train the final risk control model according to the filtered feature activation set. The use of this method can improve the prediction accuracy of the risk control model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of risk prediction, and in particular to a risk control model training method, risk prediction method, device, and computer equipment. Background Art

[0002] With the development of Internet technology, various Internet products involved in transactions on the current Internet have the characteristics of fast launch and short history. For Internet products that have been in operation for a short time, there is no sufficient operational data support. Therefore, risk control for such Internet products that have been in operation for a short time has always been a problem to be solved.

[0003] In the related technology, network models are generally used to perform risk control on Internet products. By accumulating a certain amount of Internet product data, the risk control model is modeled and trained. Among them, when the risk control model is cold-started for an Internet product that has been in operation for a short time, experienced experts generally select feature labels, and use the feature labels specified by the experts to model and train the risk control model. The risk control model is then used to predict the risks of the data generated by the user's use of the Internet product to determine whether the user is currently at risk.

[0004] The above-mentioned cold start method of the risk control model subjectively selects feature labels, resulting in a low accuracy rate of the final risk control model. Summary of the invention

[0005] Based on this, it is necessary to provide a training method, device, computer equipment, computer-readable storage medium and computer program product for a risk control model that can improve the accuracy of risk control in response to the above-mentioned technical problems.

[0006] In a first aspect, the present application provides a method for training a risk control model. The method comprises:

[0007] According to a plurality of feature labels included in the risk control training sample set and feature label values ​​corresponding to the feature labels, a feature column vector corresponding to the feature label is obtained;

[0008] Determine the correlation coefficient between the feature column vector and a preset classification identification vector, and obtain the correlation coefficient between the feature label and the classification identification vector; wherein each value in the classification identification vector is used to represent a risk control classification identification among several preset risk control classification identifications;

[0009] Among the feature labels of the risk control training samples, the feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions are screened, and the screened feature labels are added to the feature activation set;

[0010] For any risk control training sample in the risk control training sample set, the feature label values ​​of each feature label of the risk control training sample in the feature activation set are input into the constructed risk control model, and the parameters of the risk control model are adjusted according to the output of the risk control model and the classification label of the risk control training sample until a risk control model that meets the preset conditions is obtained.

[0011] In one embodiment, determining the correlation coefficient between the feature column vector and a preset classification identification vector to obtain the correlation coefficient corresponding to the feature label includes:

[0012] Determine the covariance of the feature column vector and the preset classification identification vector respectively, and obtain the correlation coefficient between the feature column vector and the preset classification identification vector;

[0013] The correlation coefficient between the feature column vector and the preset classification identification vector is used as the correlation coefficient of the feature label.

[0014] In one embodiment, the feature tags of the risk control training samples are screened for feature tags whose corresponding correlation coefficients meet preset feature screening conditions, and the screened feature tags are added to the feature activation set, including:

[0015] Sorting the correlation coefficients corresponding to the feature labels of the risk control training samples;

[0016] Among the feature labels of the risk control training samples, feature labels whose corresponding correlation coefficients rank in the front by a preset number are screened, and the screened feature labels are added to the feature activation set.

[0017] In one embodiment, the feature tags of the risk control training samples are screened for feature tags whose corresponding correlation coefficients meet preset feature screening conditions, and the screened feature tags are added to the feature activation set, including:

[0018] Sorting the correlation coefficients corresponding to the feature labels of the risk control training samples;

[0019] Select a preset number of feature labels with corresponding correlation coefficients to add to the feature activation set, and add other feature labels to the non-feature activation set;

[0020] Training an initial risk control model based on the feature activation set, and obtaining a model error corresponding to the feature activation set as a historical model error;

[0021] According to a preset exchange order, at least one feature label in the feature activation set and the non-feature activation set is exchanged to obtain a current feature activation set;

[0022] Training the initial risk model based on the current activated feature set to obtain a current model error corresponding to the current feature activation set;

[0023] When the absolute value of the difference between the current model error and the historical model error is higher than the preset error threshold, and the current model error is smaller than the historical model error, the historical model error is updated to the current error, and the step of exchanging at least one feature label in the feature activation set and the non-feature activation set in a preset exchange order to obtain the current feature activation set is returned to execute, until the absolute value of the difference between the current model error and the historical model error is smaller than the preset error threshold, and the current feature activation set is used as the final feature activation set.

[0024] In one embodiment, the feature tags of the risk control training samples are screened for feature tags whose corresponding correlation coefficients meet preset feature screening conditions, and the screened feature tags are added to the feature activation set, including:

[0025] Among the feature labels of the risk control training samples, select feature labels whose corresponding correlation coefficients are within a preset range;

[0026] The filtered feature labels are added to the feature activation set.

[0027] In one embodiment, before obtaining, for each feature label of the risk control training sample, a feature column vector corresponding to the feature label in the risk control training sample set, the method further includes:

[0028] For each risk control sample, verify each feature label value of the risk control sample;

[0029] When each feature label value of the risk control sample passes verification, the risk control sample is added to the risk control training sample set.

[0030] In one embodiment, before obtaining, for each feature label of the risk control training sample, a feature column vector corresponding to the feature label in the risk control training sample set, the method further includes:

[0031] Determine the number of positive samples marked with positive classification labels and the number of negative samples marked with negative classification labels in the risk control training sample set;

[0032] In the case where the number of negative samples is less than the number of positive samples, copying a number of risk control training samples labeled as negative classification labels in the risk control training sample set so that the number of positive samples is equal to the number of negative samples;

[0033] In the case that the number of negative samples is greater than the number of positive samples, a number of risk control training samples labeled as positive classification labels are copied in the risk control training sample set so that the number of positive samples is equal to the number of negative samples.

[0034] In a second aspect, the present application also provides a risk prediction method. The method comprises:

[0035] Obtain the risk data to be predicted corresponding to the user, determine the label value of each feature label of the user in the feature activation set, and obtain the feature vector corresponding to the risk data to be predicted; the feature activation set is the feature activation set in any of the above-mentioned risk control model training methods;

[0036] Inputting the feature vector into the risk control model in any of the risk control model training methods described above;

[0037] A risk label of the predicted risk data is obtained according to the output of the risk control model.

[0038] In a third aspect, the present application also provides a risk control model training device, the device comprising:

[0039] An acquisition module, used to acquire a feature column vector corresponding to a feature label according to a plurality of feature labels contained in a risk control training sample set and feature label values ​​corresponding to the feature labels;

[0040] A determination module, used to determine the correlation coefficient between the feature column vector and a preset classification identification vector, and obtain the correlation coefficient between the feature label and the classification identification vector; wherein each value in the classification identification vector is used to represent a risk control classification identification among several preset risk control classification identifications;

[0041] A screening module, used to screen the feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions from among the feature labels of the risk control training samples, and add the screened feature labels to the feature activation set;

[0042] A training module is used to input the feature label values ​​of each feature label of the risk control training sample in the feature activation set into the constructed risk control model for any risk control training sample in the risk control training sample set, and adjust the parameters of the risk control model according to the output of the risk control model and the classification label of the risk control training sample until a risk control model that meets preset conditions is obtained.

[0043] In a fourth aspect, the present application also provides a risk prediction device. The device comprises:

[0044] An acquisition module is used to acquire the risk data to be predicted corresponding to the user, determine the label value of each feature label of the user in the feature activation set, and obtain the feature vector corresponding to the risk data to be predicted; the feature activation set is the feature activation set in any of the above-mentioned risk control model training methods;

[0045] An input module, used for inputting the feature vector into the risk control model in any of the above-mentioned risk control model training methods;

[0046] The prediction module is used to obtain the risk label of the predicted risk data according to the output of the risk control model.

[0047] In a fifth aspect, the present application further provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0048] According to a plurality of feature labels included in the risk control training sample set and feature label values ​​corresponding to the feature labels, a feature column vector corresponding to the feature label is obtained;

[0049] Determine the correlation coefficient between the feature column vector and a preset classification identification vector, and obtain the correlation coefficient between the feature label and the classification identification vector; wherein each value in the classification identification vector is used to represent a risk control classification identification among several preset risk control classification identifications;

[0050] Among the feature labels of the risk control training samples, the feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions are screened, and the screened feature labels are added to the feature activation set;

[0051] For any risk control training sample in the risk control training sample set, the feature label values ​​of each feature label of the risk control training sample in the feature activation set are input into the constructed risk control model, and the parameters of the risk control model are adjusted according to the output of the risk control model and the classification label of the risk control training sample until a risk control model that meets the preset conditions is obtained.

[0052] In a sixth aspect, the present application further provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0053] Obtain the risk data to be predicted corresponding to the user, determine the label value of each feature label of the user in the feature activation set, and obtain the feature vector corresponding to the risk data to be predicted; the feature activation set is the feature activation set in any of the above-mentioned risk control model training methods;

[0054] Inputting the feature vector into the risk control model in any of the risk control model training methods described above;

[0055] A risk label of the predicted risk data is obtained according to the output of the risk control model.

[0056] In a seventh aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0057] According to a plurality of feature labels included in the risk control training sample set and feature label values ​​corresponding to the feature labels, a feature column vector corresponding to the feature label is obtained;

[0058] Determine the correlation coefficient between the feature column vector and a preset classification identification vector, and obtain the correlation coefficient between the feature label and the classification identification vector; wherein each value in the classification identification vector is used to represent a risk control classification identification among several preset risk control classification identifications;

[0059] Among the feature labels of the risk control training samples, the feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions are screened, and the screened feature labels are added to the feature activation set;

[0060] For any risk control training sample in the risk control training sample set, the feature label values ​​of each feature label of the risk control training sample in the feature activation set are input into the constructed risk control model, and the parameters of the risk control model are adjusted according to the output of the risk control model and the classification label of the risk control training sample until a risk control model that meets the preset conditions is obtained.

[0061] In an eighth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0062] Obtain the risk data to be predicted corresponding to the user, determine the label value of each feature label of the user in the feature activation set, and obtain the feature vector corresponding to the risk data to be predicted; the feature activation set is the feature activation set in any of the above-mentioned risk control model training methods;

[0063] Inputting the feature vector into the risk control model in any of the risk control model training methods described above;

[0064] A risk label of the predicted risk data is obtained according to the output of the risk control model.

[0065] In a ninth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0066] According to a plurality of feature labels included in the risk control training sample set and feature label values ​​corresponding to the feature labels, a feature column vector corresponding to the feature label is obtained;

[0067] Determine the correlation coefficient between the feature column vector and a preset classification identification vector, and obtain the correlation coefficient between the feature label and the classification identification vector; wherein each value in the classification identification vector is used to represent a risk control classification identification among several preset risk control classification identifications;

[0068] Among the feature labels of the risk control training samples, the feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions are screened, and the screened feature labels are added to the feature activation set;

[0069] For any risk control training sample in the risk control training sample set, the feature label values ​​of each feature label of the risk control training sample in the feature activation set are input into the constructed risk control model, and the parameters of the risk control model are adjusted according to the output of the risk control model and the classification label of the risk control training sample until a risk control model that meets the preset conditions is obtained.

[0070] In a tenth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0071] Obtain the risk data to be predicted corresponding to the user, determine the label value of each feature label of the user in the feature activation set, and obtain the feature vector corresponding to the risk data to be predicted; the feature activation set is the feature activation set in any of the above-mentioned risk control model training methods;

[0072] Inputting the feature vector into the risk control model in any of the risk control model training methods described above;

[0073] A risk label of the predicted risk data is obtained according to the output of the risk control model.

[0074] The training method, device, computer equipment, storage medium and computer program product of the above-mentioned risk control model screen out feature labels that meet preset conditions with the identification feature vector through correlation coefficients. Compared with the manually specified feature labels in the related art, the selected feature labels are more accurate and have better feature performance. Therefore, by training the risk control model with the feature labels screened by the present application, the prediction accuracy of the trained risk control model can be higher than that in the related art. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 A schematic diagram of a flow chart of a method for training a risk control model in one embodiment;

[0076] Figure 2 A schematic diagram of a process for selecting a feature activation set in one embodiment;

[0077] Figure 3 A schematic diagram of a risk prediction method in one embodiment;

[0078] Figure 4 is a structural block diagram of a training device for a risk control model in one embodiment;

[0079] Figure 5 is a structural block diagram of a risk prediction device in one embodiment;

[0080] Figure 6 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0081] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0082] With the development of Internet technology, various Internet products involved in transactions on the current Internet have the characteristics of fast launch and short history. For Internet products that have been in operation for a short time, there is no sufficient operational data support. Therefore, risk control for such Internet products that have been in operation for a short time has always been a problem to be solved.

[0083] Specifically, in the early stage of Internet products, there are fewer users and less data generated by users using Internet products, so it is difficult to obtain enough data to support enough training samples to support the training of risk control models. For example, Internet products such as third-party payment platforms, e-commerce platforms, and knowledge payment platforms, when users use these Internet products, there may be various risks, such as accounts being hacked to post inappropriate content, stolen accounts causing financial losses, etc.

[0084] In related technologies, network models are generally used to perform risk control on Internet products. By accumulating certain Internet product data, risk control models are modeled and trained. Among them, when the risk control model is cold-started for an Internet product that has been in operation for a short time, experienced experts generally select feature labels, and model and train the risk control model using feature labels specified by experts, such as common features such as account login location, account login times, account interaction activity, etc., thereby reducing the risk control model's dependence on the number of training samples required.

[0085] Obviously, the modeling and training of risk control models in the above manner will be highly subjective. For some special Internet products, the interaction methods, login methods, etc. may be unique, and the general features may not perform very well, which makes the prediction accuracy of the risk control model finally trained low.

[0086] Based on this, the present application proposes a training method for a risk control model, which obtains a feature column vector corresponding to a feature label according to multiple feature labels contained in a risk control training sample set and feature label values ​​corresponding to the feature labels. Determine the correlation coefficient between the feature column vector and the preset classification identification vector, and obtain the correlation coefficient between the feature label and the classification identification vector; wherein each value in the classification identification vector is used to represent a risk control classification identifier among several preset risk control classification identifiers. Among the feature labels of the risk control training sample, filter the feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions, and add the filtered feature labels to the feature activation set. For any risk control training sample in the risk control training sample set, input the feature label values ​​of each feature label of the risk control training sample in the feature activation set into the constructed risk control model, and adjust the parameters of the risk control model according to the output of the risk control model and the classification label of the risk control training sample until a risk control model that meets the preset conditions is obtained. .

[0087] Through the risk control model training method provided in this application, feature labels that are helpful for risk control model training are screened out through correlation coefficients, and finally the risk control model is trained with the screened feature labels, so that the accuracy of the trained risk control model is higher than that in related technologies.

[0088] In one embodiment, Figure 1 As shown, a training method for a risk control model is provided. This embodiment uses the method applied to a terminal as an example. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0089] Step 101 obtains a feature column vector corresponding to a feature label according to a plurality of feature labels included in a risk control training sample set and feature label values ​​corresponding to the feature labels.

[0090] Among them, the feature column vector corresponding to the feature label is composed of the feature label values ​​corresponding to the feature label of each risk control training sample in the risk control training sample set. The feature labels contained in the training sample set are feature labels of different dimensions of Internet products. The terminal device can traverse all the data generated by users using Internet products to obtain various data items as feature labels of the Internet products. For example, when users use the above-mentioned third-party payment platforms, e-commerce platforms, knowledge payment platforms and other Internet products, various data items may be generated, such as user name, last login time, number of logins, number of interactions, browsing preferences, etc.

[0091] Specifically, the risk control training sample set generally includes multiple risk control training samples, each risk control training sample includes feature label values ​​of multiple feature labels. For a feature label, the terminal device obtains the feature label value of each risk control training sample in the risk control training sample set for the feature label, and obtains a feature column vector composed of the feature label values ​​of each risk control training sample for the feature label.

[0092] As shown in Table 1, this is the risk control sample set shown in this application:

[0093] Table 1 Risk control sample set

[0094] Feature tag 1 Feature tag 2 Feature tag 3 Sample A Feature label value A1 Feature label value A2 Feature label value A3 Sample B Feature label value B1 Feature label value B2 Feature label value B3 Sample C Feature label value C1 Feature label value C2 Feature label value C3 Sample D Feature label value D1 Feature label value D2 Feature label value D3

[0095] Among them, the sample feature vector of sample A is: [feature label value A1, feature label value A2, feature label value A3]. The feature column vector corresponding to feature label 1 contained in the risk control training sample set is: [feature label value A1, feature label value B1, feature label value C1, feature label value D1].

[0096] In one embodiment, the terminal device stores the risk control training sample set in the form of a two-dimensional table, where each row corresponds to a risk control training sample, each column corresponds to a feature label, and the feature label value of each risk control training sample for the feature label. The terminal device obtains the feature column vector corresponding to each feature label by traversing the feature label value corresponding to the column corresponding to each feature label.

[0097] Step 103: determine the correlation coefficient between the feature column vector and the preset classification identification vector, and obtain the correlation coefficient between the feature label and the classification identification vector.

[0098] Each value in the classification identification vector is used to represent one of several preset risk control classification identifications. For example, the risk control classification results include two results, namely, risky and risk-free. The risky classification identification can be 1, and the risk-free classification identification can be 0, so the classification identification vector can be [0,1] or [1,0]. For another example, the risk control classification results include three results, namely, high risk, medium risk, and low risk. The high risk classification identification can be 3, the medium risk classification identification can be 2, and the low risk classification identification can be 1, so the classification identification vector can be [3,2,1] or [1,2,3]. The higher the absolute value of the correlation coefficient between the feature label and the classification identification vector, the higher the feature performance of the feature label for predicting the classification.

[0099] Specifically, for each feature tag, the terminal device obtains the correlation coefficient between the feature column vector and the preset classification identification vector according to the feature column vector corresponding to the feature tag and the preset classification identification vector, as the correlation coefficient corresponding to the feature tag.

[0100] Step 105: Among the feature labels of the risk control training samples, the feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions are selected, and the screened feature labels are added to the feature activation set.

[0101] Specifically, after the terminal device obtains the correlation coefficient corresponding to each feature tag, it selects the feature tag corresponding to the correlation coefficient that meets the preset feature screening condition, such as screening the feature tag whose absolute value of the correlation coefficient is at the previous preset position, or screening the feature tag whose absolute value of the correlation coefficient is within the preset range. Afterwards, the terminal device initializes the feature activation set to an empty set, traverses the correlation coefficients of each feature tag, and determines whether the correlation coefficients of each feature tag meet the preset feature screening condition. If so, the feature tag is added to the feature activation set. If not, no processing is performed. Finally, the terminal device adds the screened feature tags to the feature activation set. For example, if 10 feature tags are included, the terminal device obtains 5 feature tags after screening, and adds the 5 screened feature tags to the feature activation set.

[0102] Step 107: For any risk control training sample in the risk control training sample set, the feature label value of each feature label of the risk control training sample in the feature activation set is input into the constructed risk control model, and the parameters of the risk control model are adjusted according to the output of the risk control model and the classification label of the risk control training sample until a risk control model that meets the preset conditions is obtained.

[0103] Specifically, after the terminal device obtains the filtered feature activation set, it uses the feature activation set to train the risk control model. When the terminal device is training the model, for each risk control training sample, the feature vector composed of the feature label values ​​of each feature label in the feature activation set in the risk control training sample is used as the input of the risk control model, that is, the feature label values ​​of each feature label of the risk control training sample in the feature activation set are input into the constructed risk control model, and then the parameters of the risk control model are adjusted according to the output of the risk control model and the classification label of the risk control training sample, and the training is cyclically performed until a risk control model that meets the preset conditions is obtained, such as the error between the output of the risk control model and the label of the risk control training sample is less than the preset value, or the risk control model is cyclically trained for a preset number of times, etc.

[0104] In this embodiment, feature labels with a high correlation with the identification feature vector are screened out through the correlation coefficient. Compared with the manually specified feature labels in the related art, the feature performance is better. Therefore, by training the risk control model with the feature labels screened out by the present application, the prediction accuracy of the trained risk control model can be higher than that in the related art.

[0105] In one embodiment, the above step 103 specifically includes:

[0106] Step 103A: respectively determine the covariance between the feature column vector and the preset classification identification vector, and obtain the correlation coefficient between the feature column vector and the preset classification identification vector.

[0107] Specifically, the terminal device determines the feature column vector corresponding to the feature label, calculates the covariance of the feature column vector and the preset classification identification vector, and obtains the correlation coefficient between the feature column vector and the preset classification identification vector, as shown below, which is a correlation coefficient calculation formula shown in this application:

[0108]

[0109] Among them, U represents the feature column vector, V represents the classification identification vector, Bcov(U) is the spherical covariance of the feature column vector, Bcov(V) is the spherical covariance of the classification identification vector, and Bcor(U,V) is the correlation coefficient between the feature column vector and the preset classification identification vector.

[0110] In addition, the terminal device may also determine the covariance of the classification identification vector preset by the feature column vector and the correlation coefficient between the feature column vector and the preset classification identification vector according to other covariance calculation methods.

[0111] Step 103B: taking the correlation coefficient between the feature column vector and the preset classification identification vector as the correlation coefficient of the feature label.

[0112] Specifically, after obtaining the correlation coefficient between the feature column vector and the preset classification identification vector, the terminal device uses the obtained determined correlation coefficient as the correlation coefficient between the feature label corresponding to the feature column vector and the classification identification.

[0113] In this embodiment, the correlation coefficient between the feature label and the classification identifier is determined by the covariance between the feature column vector and the preset classification identifier vector.

[0114] In one embodiment, the above step 105 specifically includes:

[0115] Step A01: Sort the correlation coefficients corresponding to the feature labels of the risk control training samples.

[0116] Specifically, after obtaining the correlation coefficient corresponding to each feature tag, the terminal device sorts the feature tags from large to small according to the absolute value of the correlation coefficient corresponding to each feature tag. The absolute value of the correlation coefficient corresponding to the feature tag is proportional to the correlation relationship between the feature tag and the score identifier. The larger the absolute value of the correlation coefficient, the higher the correlation relationship between the feature tag and the score identifier. The absolute value of the correlation coefficient is 0, and there is no correlation relationship between the feature tag and the score identifier.

[0117] Step A02: among the feature labels of the risk control training samples, select the feature labels whose corresponding correlation coefficients rank in the front by a preset number, and add the selected feature labels to the feature activation set.

[0118] Specifically, after the terminal device sorts each feature label according to the correlation coefficient corresponding to each feature label, it selects a preset number of feature labels with corresponding correlation coefficients in the front from each feature label of the risk control training sample in order from large to small, and adds the selected feature labels to the feature activation set as the feature activation set for the final model training.

[0119] In one embodiment, the training method of the risk control model further includes:

[0120] Step A03: add other feature tags that are not added to the feature activation set to the non-feature activation set.

[0121] Step A04: Train the initial risk control model based on the feature activation set, and obtain the model error corresponding to the feature activation set as the historical model error.

[0122] Step A05: exchange at least one feature tag in the feature activation set and the non-feature activation set according to a preset exchange order to obtain a current feature activation set.

[0123] Step A06: Train the initial risk model based on the current activated feature set to obtain the current model error corresponding to the current feature activation set.

[0124] Step A07: When the absolute value of the difference between the current model error and the historical model error is higher than the preset error threshold and the current model error is smaller than the historical model error, update the historical model error to the current error and return to execute step A05 until the absolute value of the difference between the current model error and the historical model error is smaller than the preset error threshold, and use the current feature activation set as the final feature activation set.

[0125] In this embodiment, feature labels with high absolute values ​​of correlation coefficients are selected for model training.

[0126] In one embodiment, the above step 105 specifically includes:

[0127] Step B01: sort the correlation coefficients corresponding to each feature label.

[0128] This step can refer to the description of step A1, and will not be described in detail here.

[0129] Step B02: Select a preset number of feature tags with corresponding correlation coefficients and add them to the feature activation set, and add other tags to the non-feature activation set.

[0130] Specifically, the terminal device initializes the feature activation set as an empty set, and then sorts each feature tag according to the correlation coefficient corresponding to each feature tag, and then selects a preset number of feature tags with the corresponding correlation coefficient in the order from large to small, and adds the selected feature tags to the feature activation set to complete the update of the feature activation set. The terminal device initializes the non-feature activation set as an empty set, selects a preset number of feature tags to add to the feature activation set, and then adds the remaining feature tags to the non-feature activation set to complete the update of the non-feature activation set. For example, the risk control training sample set includes 15 feature tags. The device adds the top ten feature tags to the feature activation set, and the bottom five feature tags to the non-feature activation set.

[0131] Step B03: Train the initial risk control model based on the feature activation set, and obtain the model error corresponding to the feature activation set as the historical model error.

[0132] Specifically, the terminal device trains the initial risk control model based on the current feature activation set. When the terminal device performs model training, for each risk control training sample, the feature vector composed of the feature label values ​​of each feature label in the current feature activation set in the risk control training sample is used as the input of the initial risk control model, and then the model error of the initial risk control model trained based on the current feature activation set is obtained according to the output of the initial risk control model and the error between the classification labels of the risk control training sample, and the obtained model error is used as the historical model error.

[0133] Step B04: exchange at least one feature tag in the feature activation set and the non-feature activation set according to a preset exchange order to obtain a current feature activation set.

[0134] Specifically, in order to obtain a feature activation set with better feature performance, at least one feature tag in the feature activation set and the non-feature activation set can be exchanged according to a preset exchange strategy, for example, the feature tags ranked in the last X positions in the feature activation set are exchanged with the feature tags ranked in the first X positions in the non-feature activation set, or at least one feature tag is exchanged among the feature tags ranked in the last Y positions in the feature activation set and the feature tags ranked in the first Y positions in the non-feature activation set. Among them, the number of feature tags contained in the feature activation set before the exchange and the feature activation set after the exchange remains unchanged, that is, the exchange of the feature activation set and the non-feature activation set is a one-to-one exchange. After the terminal device exchanges at least one feature tag in the feature activation set and the non-feature activation set, the current feature activation set is obtained.

[0135] It should be noted that no matter which exchange order is used, the feature activation set after the exchange is different from the historical feature activation set before the exchange. For example, before the initial exchange, the feature activation set is P0, after the first exchange, the feature activation set is P1, after the second exchange, the feature activation set is P2, and so on, until a qualified feature activation set is obtained. The feature activation sets P0, P1, P2... are different from each other.

[0136] Step B05: Train the initial risk model based on the current activated feature set to obtain the current model error corresponding to the current feature activation set.

[0137] Specifically, after the terminal device exchanges at least one label in the feature activation set and the non-feature activation set, it obtains the current feature activation set, and then trains the initial risk model based on the current activated feature set to obtain the current model error corresponding to the current feature activation set. When the terminal device performs model training, for each risk control training sample, the feature vector composed of the feature label values ​​of each feature label in the current feature activation set in the risk control training sample is used as the input of the initial risk control model, and then the model error of the initial risk control model trained based on the current feature activation set is obtained based on the output of the initial risk control model and the error between the classification labels of the risk control training sample, and the obtained model error is used as the current model error, that is, the current model error corresponding to the current feature activation set.

[0138] Step B06: When the absolute value of the difference between the current model error and the historical model error is higher than the preset error threshold and the current model error is smaller than the historical model error, update the historical model error to the current error and return to step B04 until the absolute value of the difference between the current model error and the historical model error is smaller than the preset error threshold, and use the current feature activation set as the final feature activation set.

[0139] Specifically, after the terminal device obtains the current model error corresponding to the current feature activation set, if it is determined that the absolute value of the difference between the current model error and the historical model error is higher than the preset error threshold, that is, the current model error is much smaller than the historical model error, then it is determined that the feature performance of the feature activation set after the exchange is significantly better than the feature activation set before the exchange, the historical model error is updated, and the current model error corresponding to the current feature activation set is used as the historical model error, and then return to execute step B04.

[0140] In one embodiment, after the terminal device obtains the current model error corresponding to the current feature activation set, if it is determined that the absolute value of the difference between the current model error and the historical model error is less than a preset error threshold, and the current model error is less than the historical model error, then it is determined that the feature performance of the feature activation set after the exchange is not significantly better than the feature performance of the feature activation set before the exchange, the update of the feature activation set is completed, and the loop is exited, and the current feature activation set is used as the final feature activation set.

[0141] After the terminal device obtains the current model error corresponding to the current feature activation set, if it is determined that the absolute value of the difference between the current model error and the historical model error is less than the preset error threshold, and the current model error is greater than the historical model error, it returns to execute step B04.

[0142] In one embodiment, when the terminal device determines that the current model error is not less than the historical model error, it returns to execute step B04. Specifically, after the terminal device obtains the current model error corresponding to the current feature activation set, if the current model error is greater than the historical model error, it is determined that the feature performance of the feature activation set after the exchange is inferior to the feature activation set before the exchange. At this time, there is no need to update the historical model error, and the terminal device directly returns to execute step B04.

[0143] In one embodiment, after the terminal device obtains the current model error corresponding to the current feature activation set, if the current model error is greater than the historical model error, then it is determined that the feature performance of the feature activation set after the exchange is inferior to the feature activation set before the exchange. At this time, there is no need to update the historical model error. The current feature activation set is restored to the feature activation set before the exchange and then returns to execute step B04. For example, the current feature activation set is {A1, A2, A4}, and the current feature activation set before this cyclic exchange is {A1, A2, A3}, then the feature activation set is restored to {A1, A2, A3} and then returns to execute step B04.

[0144] like Figure 2 , which is a flow chart of the execution process of a feature activation set according to an embodiment of the present application.

[0145] Step 201: After the terminal device sorts each feature tag according to the correlation coefficient corresponding to each feature tag, it selects a preset number of feature tags with corresponding correlation coefficients in the order from large to small to add them to the feature activation set, and adds the remaining feature tags to the non-feature activation set, completing the initialization of the feature activation set and the non-feature activation set. The terminal device trains the initial risk control model according to the initially obtained feature activation set, obtains the model error corresponding to the initial feature activation set, and initializes the historical model error to the model error corresponding to the initial feature activation set.

[0146] Step 202: The terminal device exchanges at least one feature tag in a feature activation set and a non-feature activation set according to a preset exchange order.

[0147] Step 203: The terminal device trains an initial risk control model based on the current feature activation set to obtain a current model error corresponding to the current feature activation set.

[0148] Step 204 , the terminal device determines whether the current model error is less than the historical model error. If so, continue to execute step 205 ; if not, return to execute step 202 .

[0149] Step 205 , the terminal device determines whether the absolute value of the difference between the current model error and the historical model error is less than a preset error threshold. If so, execute step 206 ; if not, exit the loop and execute step 207 .

[0150] Step 206 : The terminal device updates the historical model error, takes the current error model error as the historical model error, and then returns to execute step 202 .

[0151] Step 207: The terminal device uses the current feature activation set as the final feature activation set, that is, as the feature activation set for risk control model training.

[0152] In this embodiment, a feature activation set with the best feature performance is obtained through multiple iterations, and the risk control model trained based on the feature activation set has a higher prediction accuracy.

[0153] In one embodiment, the above step 105 specifically includes:

[0154] Step C1: Among the feature labels in the risk control training sample set, select the feature labels whose corresponding correlation coefficients are within a preset range.

[0155] Specifically, the technical personnel can pre-set the range of correlation coefficients in the terminal device according to actual needs. After the terminal device obtains the correlation coefficients corresponding to each feature label of the risk control training sample set, it traverses the correlation coefficients corresponding to each feature label to determine whether the correlation coefficient corresponding to the feature label is within the preset range, and then screens out the feature labels whose corresponding correlation coefficients are within the preset range.

[0156] Step C2: Add the filtered feature labels to the feature activation set.

[0157] Specifically, after the terminal device screens out feature tags with corresponding correlation coefficients within a preset range, the screened feature tags are added to a feature activation set, and the feature tags in the feature activation set are used as feature tags for training the risk control model.

[0158] In this embodiment, feature tags with correlation coefficients within a preset range are selected, and the screening range can be adjusted according to the number of feature tags actually screened out, so that the number of feature tags screened out is appropriate and the correlation coefficients with the classification identification vector are appropriate.

[0159] In one embodiment, before executing step 101, the method further includes:

[0160] Step 109: For each risk control sample, verify each feature label value of the risk control sample.

[0161] Specifically, after the terminal device obtains multiple risk control samples, there will be dirty samples in the samples, for example, some feature labels of some risk control samples are missing feature label values, such as feature label values ​​of NA, NULL, null, none, etc., or feature label values ​​of some feature labels of some risk control samples are obviously abnormal, such as feature label values ​​that exceed a certain range or deviate from the normal distribution, or feature label values ​​of some feature labels of some risk control samples are obviously wrong, such as data that does not meet the requirements and range standards, such as the age range is 0-120, and data exceeding 120 is wrong data. For each risk control sample, the terminal device verifies each feature label value of the risk control sample in turn according to the verification requirements corresponding to each feature label, and determines whether each feature label value of the risk control sample passes the verification.

[0162] In one embodiment, after the terminal device obtains all the data generated by each user when using the Internet product, for each user, the user identifier of the user is used as the unique identifier of the risk control sample, and each data item generated by the user when using the Internet product is obtained. Each data item is a feature label to form a risk control sample corresponding to the user.

[0163] Step 111: When each feature label value of the risk control sample passes verification, the risk control sample is added to the risk control training sample set.

[0164] Specifically, after the terminal device verifies each risk control sample, it adds the risk control sample that has passed the verification of each feature label value to the risk control training sample as one of the risk control training samples for training the risk control model.

[0165] In this embodiment, before training, data cleaning is performed on the risk control samples to obtain a high-quality risk control training sample set.

[0166] In one embodiment, before executing step 101, the method further includes:

[0167] Step 113: Determine a first number of positive samples marked with positive classification labels and a second number of negative samples marked with negative classification labels in the risk control training sample set.

[0168] Among them, the positive classification label refers to the risk-free risk control training sample, and the negative classification label refers to the risky risk control training sample. At this time, the risk control model is a binary classification model, which is used to classify the input data into one of the categories.

[0169] Specifically, the terminal device traverses the classification label of each risk control training sample in the risk control training sample set, initializes the first number and the second number to 0, and increases the first number corresponding to the positive sample by one each time a risk control training sample marked with a positive classification label is traversed, and increases the second number corresponding to the negative sample by one each time a risk control training sample marked with a negative classification label is traversed, until each risk control training sample in the risk control training sample set is traversed, and the first number of positive samples marked with positive classification labels and the second number of negative samples marked with negative classification labels in the risk control training sample set are obtained.

[0170] Step 115: When the second number is smaller than the first number, copy a number of risk control training samples labeled as negative classification labels in the risk control training sample set so that the second number is equal to the first number.

[0171] Specifically, the terminal device determines that the second number corresponding to the negative samples is smaller than the first number corresponding to the positive samples, indicating that the number of positive and negative samples is unbalanced. At this time, several risk control training samples labeled with negative classification labels can be copied in the risk control training sample set to make the second number equal to the first number. In this way, the same risk control training sample labeled with a negative classification label will be repeated at least once for model training, thereby reducing the deviation in the training of the risk control model caused by the second number corresponding to the negative samples being less than the first number corresponding to the positive samples.

[0172] Step 117: When the second number is greater than the first number, copy a number of risk control training samples labeled as positive classification labels in the risk control training sample set so that the second number is equal to the first number.

[0173] Specifically, the terminal device determines that the first number is less than the second number, indicating that the number of positive and negative samples is unbalanced. At this time, several risk control training samples labeled as positive classification labels can be copied in the risk control training sample set to make the second number equal to the first number. In this way, the same risk control training sample labeled as positive classification label will be repeated at least once for model training, thereby reducing the deviation in the training of the risk control model caused by the fact that the first number corresponding to the positive samples is less than the second number corresponding to the negative samples.

[0174] In this embodiment, the model error caused by sample imbalance is reduced by balancing the number of positive and negative samples.

[0175] In one embodiment, Figure 3 As shown, a risk prediction method is provided. This embodiment uses the method applied to a terminal as an example for illustration. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0176] Step 301: Obtain the risk data to be predicted corresponding to the user, determine the label value of each feature label of the user in the feature activation set, and obtain the feature vector corresponding to the risk data to be predicted.

[0177] Among them, the feature activation set is the feature activation set finally obtained by the training method of the above-mentioned risk control model.

[0178] Specifically, when users use Internet application products, corresponding data will be generated. The terminal device collects the data and obtains the corresponding risk data to be predicted for the user. It then extracts features from the risk data to be predicted, determines the label values ​​of each feature label of the user in the feature activation set, and then obtains the feature vector corresponding to the risk data to be predicted.

[0179] Step 303: Input the feature vector into the trained risk control model.

[0180] Among them, the trained risk control model is the trained risk control model described in any one of the above-mentioned risk control model training methods.

[0181] Specifically, after the terminal device obtains the feature vector corresponding to the risk data to be predicted by the user, it inputs the feature vector into the trained risk control model, so that the risk control model predicts the risk of the user according to the feature vector.

[0182] Step 305: Obtain the risk label of the risk data to be predicted according to the output of the risk control model.

[0183] Specifically, after the terminal device inputs the feature vector into the trained risk control model, the risk control model processes the feature vector and outputs the corresponding risk prediction result. According to the output of the risk control model, the risk label of the risk data to be predicted is obtained, such as having risk or no risk.

[0184] Next, a specific embodiment provided by the present application will be described in detail.

[0185] Step 1: The terminal device obtains the sample data, verifies the feature label values of each feature label of the sample data, and filters to obtain a qualified risk control training sample set, completing the cleaning of the sample data.

[0186] Step 2: The terminal device determines the number N of positive samples marked as positive samples in the risk control training sample set pos , and the number N of negative samples marked as negative samples neg .

[0187] Step 3: The terminal device calculates the small ball correlation coefficient between the feature column vector of each feature and the classification identification vector, and directly deletes the features with a ball correlation coefficient of 0.

[0188] Step 4: The terminal device sorts the remaining features according to the absolute value of the small ball correlation coefficient between their feature column vectors and the classification identification vector.

[0189] Step 5: The terminal device determines the hyperparameter P as the number of activation set features, determines the hyperparameter Q as the number of non-activation sets, and the number S of features in the exchange set, where S < P, Q, and P + Q is equal to the total number of features.

[0190] Step 6: The terminal device takes the first P of the features sorted in Step 4 as the activation set, and takes the Q features after P as the non-activation set.

[0191] Step 7: If the terminal device determines N pos < N neg , then all positive risk control training samples are weighted by N neg / N pos Vice versa.

[0192] Step 8: The terminal device uses the set P to train the initial risk control model, and obtains the initial risk control model error as α0.

[0193] Step 9: The terminal device uses the last S features in set P and the first S features in set Q to calculate any β i ∈P and j i ∈Q, exchange it, and make the new activation set P i Retrain the new classification model and calculate the model error under this model as α i , if the model training error α after exchange i <α i-1 It is believed that β i The effect is not as good as j i , then the feature β in the activation set P i Replace with j i .

[0194] Step 10: The terminal device repeats step 9 and updates the activation set until each training error α i -α i-1 <τ s , generally can be τ s =0.01slog(p)log(logn) / n.

[0195] Step 11: After the terminal device determines that the feature set P has converged, the active set at this time is recorded as P best , use this feature set to train the risk control model, and use the finally trained risk control model for risk control.

[0196] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0197] Based on the same inventive concept, the embodiment of the present application also provides a training device for a risk control model for implementing the training method for the risk control model involved above, and a risk prediction device for the risk prediction method involved. The implementation solutions to the problems provided by the training device for the risk control model and the risk prediction device are similar to the implementation solutions recorded in the training method for the risk control model and the risk prediction method mentioned above. Therefore, the specific limitations in the embodiments of the training device for one or more risk control models and the risk prediction device provided below can refer to the limitations on the training method for the risk control model and the risk prediction method above, and will not be repeated here.

[0198] In one embodiment, Figure 4 As shown, a training device for a risk control model is provided, comprising:

[0199] The acquisition module 401 is used to acquire a feature column vector corresponding to a feature label according to a plurality of feature labels included in the risk control training sample set and feature label values ​​corresponding to the feature labels;

[0200] The determination module 403 is used to determine the correlation coefficient between the feature column vector and a preset classification identification vector, and obtain the correlation coefficient between the feature label and the classification identification vector; wherein each value in the classification identification vector is used to represent a risk control classification identification among several preset risk control classification identifications;

[0201] A screening module 405 is used to screen the feature labels of the risk control training samples, the feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions, and add the screened feature labels to the feature activation set;

[0202] The training module 407 is used to input the feature label values ​​of each feature label of the risk control training sample in the feature activation set into the constructed risk control model for any risk control training sample in the risk control training sample set, and adjust the parameters of the risk control model according to the output of the risk control model and the classification label of the risk control training sample until a risk control model that meets the preset conditions is obtained.

[0203] In one embodiment, the determination module 403 specifically includes:

[0204] A covariance determination unit 403A (not shown in the figure), used to respectively determine the covariance between the feature column vector and the preset classification identification vector, and obtain the correlation coefficient between the feature column vector and the preset classification identification vector;

[0205] The correlation coefficient determination unit 403B (not shown in the figure) is used to use the correlation coefficient between the feature column vector and the preset classification identification vector as the correlation coefficient of the feature label.

[0206] In one embodiment, the screening module 405 specifically includes:

[0207] A sorting unit A1 (not shown in the figure) is used to sort the correlation coefficients corresponding to the feature tags;

[0208] The screening unit A2 (not shown in the figure) is used to screen the feature labels whose corresponding correlation coefficients rank in the front of a preset number from among the feature labels of the risk control training samples, and add the screened feature labels to the feature activation set.

[0209] In one embodiment, the screening module 405 specifically includes:

[0210] A sorting unit B1 (not shown in the figure) is used to sort the correlation coefficients corresponding to the feature tags;

[0211] A selection unit B2 (not shown in the figure) is used to select a preset number of feature tags with corresponding correlation coefficients to add to the feature activation set, and add other feature tags to the non-feature activation set;

[0212] Initialization unit B3 (not shown in the figure), used to train the initial risk control model based on the feature activation set, and obtain the model error corresponding to the feature activation set as the historical model error;

[0213] An exchange unit B4 (not shown in the figure), configured to exchange at least one feature label in the feature activation set and the non-feature activation set according to a preset exchange order to obtain a current feature activation set;

[0214] An error determination unit B5 (not shown in the figure), used to train the initial risk model based on the current activated feature set to obtain a current model error corresponding to the current feature activation set;

[0215] An updating unit B6 (not shown in the figure) is used for updating the historical model error to the current error when the absolute value of the difference between the current model error and the historical model error is higher than a preset error threshold, and returning to execute the step of exchanging at least one feature label in the feature activation set and the non-feature activation set in a preset exchange order to obtain a current feature activation set, until the absolute value of the difference between the current model error and the historical model error is less than the preset error threshold, and taking the current feature activation set as the final feature activation set;

[0216] Return unit B7 (not shown in the figure) is used to return to the step of exchanging at least one feature label in the feature activation set and the non-feature activation set in a preset exchange order to obtain the current feature activation set when it is determined that the current model error is not less than the historical model error.

[0217] In one embodiment, the screening module 405 specifically includes:

[0218] A screening unit C1 (not shown in the figure) is used to screen the feature labels whose corresponding correlation coefficients are within a preset range from among the feature labels of the risk control training samples;

[0219] The adding unit C2 (not shown in the figure) is used to add the filtered feature labels to the feature activation set.

[0220] In one embodiment, the above device further comprises:

[0221] Verification module 409 (not shown in the figure), used for verifying each characteristic label value of each risk control sample;

[0222] The adding module 411 (not shown in the figure) is used to add the risk control sample to the risk control training sample set when each feature label value of the risk control sample has passed the verification.

[0223] In one embodiment, the above device further comprises:

[0224] A sample quantity determination module 413 (not shown in the figure), used to determine a first quantity of positive samples marked with positive classification labels and a second quantity of negative samples marked with negative classification labels in the risk control training sample set;

[0225] A negative sample copying module 415 (not shown in the figure) is used to copy a number of risk control training samples labeled as negative classification labels in the risk control training sample set when the second number is less than the first number, so that the second number is equal to the first number;

[0226] The positive sample copying module 417 (not shown in the figure) is used to copy a number of risk control training samples labeled as positive classification labels in the risk control training sample set when the second number is greater than the first number so that the second number is equal to the first number.

[0227] In one embodiment, Figure 5 As shown, a risk prediction device is provided, comprising:

[0228] The acquisition module 501 is used to acquire the risk data to be predicted corresponding to the user, determine the label value of each feature label of the user in the feature activation set, and obtain the feature vector corresponding to the risk data to be predicted; the feature activation set is the feature activation set in any of the above-mentioned risk control model training methods;

[0229] An input module 503, used to input the feature vector into the risk control model in any of the above-mentioned risk control model training methods;

[0230] The prediction module 505 is used to obtain the risk label of the predicted risk data according to the output of the risk control model.

[0231] Each module in the training device and risk prediction device of the risk control model can be implemented in whole or in part by software, hardware and their combination. Each module can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each module above.

[0232] In one embodiment, a computer device is provided. The computer device may be a terminal device or a server. The internal structure diagram thereof may be as follows: Figure 6 As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless method can be implemented through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a training method or a risk prediction method for a risk control model is implemented.

[0233] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0234] In one embodiment, a computer device is also provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps in the above-mentioned training method embodiments of the risk control models or the steps in the above-mentioned risk prediction method embodiments are implemented.

[0235] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned training method embodiments of the risk control model or the steps in the above-mentioned risk prediction method embodiments are implemented.

[0236] In one embodiment, a computer program product is provided, including a computer program, which, when executed by a processor, implements the steps in the above-mentioned training method embodiments of the risk control model or the steps in the above-mentioned risk prediction method embodiments.

[0237] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0238] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0239] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0240] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for training a risk control model, characterized in that: The method comprises: According to the multiple feature tags contained in the risk control training sample set and the feature tag values ​​corresponding to the feature tags, the feature column vector corresponding to the feature tags is obtained; the feature tags contained in the risk control training sample set are feature tags of different dimensions of Internet products; the feature tags include user name, last login time, number of logins, number of interactions and browsing preferences; Determine the correlation coefficient between the feature column vector and a preset classification identification vector, and obtain the correlation coefficient between the feature label and the classification identification vector; wherein each value in the classification identification vector is used to represent a risk control classification identification among several preset risk control classification identifications; the higher the absolute value of the correlation coefficient between the feature label and the classification identification vector, the higher the characteristic performance of the feature label for predicting the classification; Among the feature labels of the risk control training samples, the feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions are screened, and the screened feature labels are added to the feature activation set; The feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions are screened out from the feature labels of the risk control training samples, and the screened feature labels are added to the feature activation set, including: Sorting the correlation coefficients corresponding to the feature labels; Select a preset number of feature labels with corresponding correlation coefficients to add to the feature activation set, and add other feature labels to the non-feature activation set; Training an initial risk control model based on the feature activation set, and obtaining a model error corresponding to the feature activation set as a historical model error; According to a preset exchange order, at least one feature tag in the feature activation set and the non-feature activation set is exchanged to obtain a current feature activation set; the number of feature tags included in the current feature activation set and the feature activation set remains unchanged, and the exchange between the feature activation set and the non-feature activation set is a one-to-one exchange; Training the initial risk control model based on the current feature activation set to obtain a current model error corresponding to the current feature activation set; When the absolute value of the difference between the current model error and the historical model error is higher than a preset error threshold, and the current model error is smaller than the historical model error, the historical model error is updated to the current model error, and the step of exchanging at least one feature label in the feature activation set and the non-feature activation set in a preset exchange order to obtain a current feature activation set is returned to the step, until the absolute value of the difference between the current model error and the historical model error is smaller than the preset error threshold, and the current feature activation set is used as the final feature activation set; For any risk control training sample in the risk control training sample set, the feature label values ​​of each feature label of the risk control training sample in the feature activation set are input into the constructed risk control model, and the parameters of the risk control model are adjusted according to the output of the risk control model and the classification label of the risk control training sample until a risk control model that meets the preset conditions is obtained.

2. The method according to claim 1, characterized in that The determining of the correlation coefficient between the feature column vector and a preset classification identification vector, and obtaining the correlation coefficient between the feature label and the classification identification vector, includes: Determine the covariance between the feature column vector and each preset classification identification vector respectively, and obtain the correlation coefficient between the feature column vector and the preset classification identification vector; The correlation coefficient between the feature column vector and the preset classification identification vector is used as the correlation coefficient between the feature label and the classification identification vector.

3. The method according to claim 1, characterized in that The feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions are screened out from the feature labels of the risk control training samples, and the screened feature labels are added to the feature activation set, including: Sorting the correlation coefficients corresponding to the feature labels; Among the feature labels of the risk control training samples, feature labels whose corresponding correlation coefficients rank in the front by a preset number are screened, and the screened feature labels are added to the feature activation set.

4. The method according to claim 1, characterized in that The feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions are screened out from the feature labels of the risk control training samples, and the screened feature labels are added to the feature activation set, including: Among the feature labels of the risk control training samples, select feature labels whose corresponding correlation coefficients are within a preset range; The filtered feature labels are added to the feature activation set.

5. The method according to claim 1, characterized in that: Before obtaining the feature column vector corresponding to the feature label according to the plurality of feature labels included in the risk control training sample set and the feature label values ​​corresponding to the feature labels, the method further includes: For each risk control sample, verify each feature label value of the risk control sample; When each feature label value of the risk control sample passes verification, the risk control sample is added to the risk control training sample set.

6. The method according to claim 5, characterized in that For each risk control sample, verifying each feature label value of the risk control sample includes: For each user, use the user ID of the user as the unique identifier of the risk control sample, obtain each data item generated by the user using the Internet product, and determine each data item as a feature label to obtain the risk control sample corresponding to the user; For each of the risk control samples, according to the verification requirements corresponding to each of the feature labels, each feature label value of the risk control sample is verified in turn to determine whether each feature label value of the risk control sample passes the verification.

7. The method according to claim 1, characterized in that Before obtaining the feature column vector corresponding to the feature label according to the plurality of feature labels included in the risk control training sample set and the feature label values ​​corresponding to the feature labels, the method further includes: Determine a first number of positive samples marked with positive classification labels and a second number of negative samples marked with negative classification labels in the risk control training sample set; When the second number is less than the first number, copying a number of risk control training samples labeled as negative classification labels in the risk control training sample set, so that the second number is equal to the first number; When the second number is greater than the first number, a number of risk control training samples labeled as positive classification labels are copied from the risk control training sample set to make the second number equal to the first number.

8. A risk prediction method, characterized in that: The method comprises: Obtain the risk data to be predicted corresponding to the user, determine the label value of each feature label of the user in the feature activation set, and obtain the feature vector corresponding to the risk data to be predicted; the feature activation set is the feature activation set in any one of the methods described in claims 1-7; Inputting the feature vector into the risk control model as described in any one of claims 1 to 7; A risk label of the predicted risk data is obtained according to the output of the risk control model.

9. A training device for a risk control model, characterized in that: The device comprises: An acquisition module is used to acquire a feature column vector corresponding to a feature tag according to a plurality of feature tags contained in a risk control training sample set and feature tag values ​​corresponding to the feature tags; the feature tags contained in the risk control training sample set are feature tags of different dimensions of Internet products; the feature tags include user name, last login time, number of logins, number of interactions and browsing preferences; A determination module, used to determine the correlation coefficient between the feature column vector and a preset classification identification vector, and obtain the correlation coefficient between the feature label and the classification identification vector; wherein each value in the classification identification vector is used to represent a risk control classification identification among several preset risk control classification identifications; A screening module is used to screen the feature labels corresponding to the correlation coefficients that meet the preset feature screening conditions from the feature labels of the risk control training samples, and add the screened feature labels to the feature activation set; the higher the absolute value of the correlation coefficient between the feature label and the classification identification vector, the higher the feature performance of the feature label for predicting the classification; The screening module is specifically used to sort the correlation coefficients corresponding to each of the feature tags; select a preset number of feature tags with corresponding correlation coefficients in the front to add to the feature activation set, and add other feature tags to the non-feature activation set; train the initial risk control model based on the feature activation set, and obtain the model error corresponding to the feature activation set as the historical model error; exchange at least one feature tag in the feature activation set and the non-feature activation set in a preset exchange order to obtain the current feature activation set; the number of feature tags contained in the current feature activation set and the feature activation set remains unchanged, and the exchange between the feature activation set and the non-feature activation set is a one-to-one exchange; based on the current The initial risk control model is trained using a feature activation set to obtain a current model error corresponding to the current feature activation set; when the absolute value of the difference between the current model error and the historical model error is higher than a preset error threshold, and the current model error is smaller than the historical model error, the historical model error is updated to be the current model error, and the step of exchanging at least one feature label in the feature activation set and the non-feature activation set in a preset exchange order to obtain the current feature activation set is returned to be executed, until the absolute value of the difference between the current model error and the historical model error is smaller than the preset error threshold, and the current feature activation set is used as the final feature activation set; A training module is used to input the feature label values ​​of each feature label of the risk control training sample in the feature activation set into the constructed risk control model for any risk control training sample in the risk control training sample set, and adjust the parameters of the risk control model according to the output of the risk control model and the classification label of the risk control training sample until a risk control model that meets preset conditions is obtained.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 or 8 are implemented.